Add WAM paper: World Pilot: Steering Vision-Language-Action Models with World-Action Priors

[279object]

Paper

• Title: World Pilot: Steering Vision-Language-Action Models with World-Action Priors
• Short name: World Pilot
• arXiv ID: 2606.12403v1
• Paper: https://arxiv.org/pdf/2606.12403
• Authors: Zefu Lin, Rongxu Cui, Junjia Xu, Xiaojuan Jin, Wenling Li, Lue Fan, Zhaoxiang Zhang
• Published: 2026-06-10
• arXiv categories: cs.RO
• Matched keywords: WAM, world-action model

README Entry

- **World Pilot**: "World Pilot: Steering Vision-Language-Action Models with World-Action Priors", arXiv 2026. ![](https://img.shields.io/badge/Implicit-ea580c)
  [[📄 Paper](https://arxiv.org/pdf/2606.12403)]

Robot Decision

• Decision: accept
• Confidence: high
• Taxonomy: cascaded_implicit_latent
• Taxonomy confidence: high
• README heading: ### Cascaded World-Action-Model
• Badges: https://img.shields.io/badge/Implicit-ea580c

Reason

The paper explicitly uses a World-Action Model to provide action-conditioned scene evolution and trajectory priors for robot manipulation policies.

Evidence

• augments VLA with priors from a World-Action Model
• jointly predicts a scene-evolution latent and coarse action-trajectory hypothesis
• evaluated on robot manipulation benchmarks and real-robot tasks

Taxonomy Evidence

• Latent Steering conditions the perception layer on a scene-evolution latent
• Action Steering supplies an anticipated trajectory as a motion prior to the action generator
• route the latent rather than a decoded future image
• WAM is kept frozen, with gradient updates restricted to the VLA parameters

Human Review Checklist

• This paper belongs in the WAM survey
• README section is correct
• Badges are correct
• Short name is correct
• Paper link is correct
• Merge this PR if accepted